AgentHarm Benchmark

What is AgentHarm?

AgentHarm tests the robustness of LLMs to jailbreak attacks. It includes a diverse set of 110 explicitly malicious agent tasks (440 with augmentations)…

Released

2024-10-11

Evaluates

安全, Safety, 智能体, Agent, 任务执行, Task Execution, 安全对齐, Safety Alignment, 不支持, Unsupported

Openness

restricted

Importer

OpenCompass Hub

Reported scores

0

Source provenance

• Original evidence https://hub.opencompass.org.cn/dataset-detail/AgentHarm

AgentHarm paper, code and dataset

• Paper https://arxiv.org/abs/2404.02151
• Paper https://arxiv.org/abs/2410.09024
• Code https://github.com/UKGovernmentBEIS/inspect_evals
• Dataset https://huggingface.co/datasets/ai-safety-institute/AgentHarm

Which sources cite AgentHarm?

• AgentHarm

Related 安全, Safety, 智能体, Agent, 任务执行, Task Execution, 安全对齐, Safety Alignment, 不支持, Unsupported benchmarks

• A-Bench
• A-OKVQA
• AbsPyramid
• ActBench
• ActionAtlas
• Ada-LEval
• AECBench
• AFQMC